Skip to content

spec : update speculative-simple #26904

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ggerganov merged 3 commits into from
Aug 11, 2026
Merged

spec : update speculative-simple #26904

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
89653b8
spec : update speculative-simple
ggerganov Aug 11, 2026
948d90a
cont : simplify
ggerganov Aug 11, 2026
c4c7f3a
cont : clean-up
ggerganov Aug 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
80 changes: 8 additions & 72 deletions common/speculative.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,12 +171,6 @@ struct common_speculative_impl {
// (optional) serialize/restore per-seq internal state (e.g. eagle3's deferred boundary).
virtual bool get_state(llama_seq_id /*seq_id*/, std::vector<uint8_t> & /*data*/) const { return false; }
virtual void set_state(llama_seq_id /*seq_id*/, const std::vector<uint8_t> & /*data*/) {}

// true if this implementation requires the target context to extract post-norm embeddings
virtual bool need_embd() const = 0;

// true if this implementation requires the target context to extract pre-norm embeddings
virtual bool need_embd_nextn() const { return false; }
};

struct common_speculative_impl_draft_simple : public common_speculative_impl {
Expand All @@ -193,6 +187,10 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
auto * ctx_dft = this->params.ctx_dft;
auto * ctx_tgt = this->params.ctx_tgt;

if (!ctx_dft) {
throw std::runtime_error("draft-simple requires a draft context");
}

SPC_TRC("%s", "adding speculative implementation 'draft-simple'\n");
SPC_TRC("- n_max=%d, n_min=%d, p_min=%f\n", this->params.n_max, this->params.n_min, this->params.p_min);
SPC_TRC("- gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n",
Expand Down Expand Up @@ -385,10 +383,6 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
// noop
}

bool need_embd() const override {
return false;
}
};


Expand Down Expand Up @@ -907,10 +901,6 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
pending_g_last[seq_id].resize(n_embd_dec);
std::memcpy(pending_g_last[seq_id].data(), data.data() + sizeof(llama_pos), (size_t) n_embd_dec * sizeof(float));
}

bool need_embd() const override {
return false;
}
};

// DFlash: block-diffusion drafting with a draft-side KV cache injection
Expand Down Expand Up @@ -1247,10 +1237,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
// noop
}

bool need_embd() const override {
return false;
}
};

struct common_speculative_impl_draft_mtp : public common_speculative_impl {
Expand Down Expand Up @@ -1689,14 +1675,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
const size_t row_bytes = (size_t) n_embd * sizeof(float);
std::memcpy(pending_h[seq_id].data(), verify_h[seq_id].data() + (size_t) i_h * n_embd, row_bytes);
}

bool need_embd() const override {
return false;
}

bool need_embd_nextn() const override {
return true;
}
};

// state of self-speculation (simple implementation, not ngram-map)
Expand Down Expand Up @@ -1743,10 +1721,6 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
// noop
}

bool need_embd() const override {
return false;
}
};

struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
Expand Down Expand Up @@ -1801,10 +1775,6 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {

common_ngram_map_accept(config[seq_id], n_accepted);
}

bool need_embd() const override {
return false;
}
};

struct common_speculative_impl_ngram_mod : public common_speculative_impl {
Expand Down Expand Up @@ -1980,10 +1950,6 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
}
}
}

bool need_embd() const override {
return false;
}
};

struct common_speculative_impl_ngram_cache : public common_speculative_impl {
Expand Down Expand Up @@ -2123,10 +2089,6 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
// noop
}

bool need_embd() const override {
return false;
}
};

struct common_speculative {
Expand Down Expand Up @@ -2322,7 +2284,6 @@ common_speculative_init_result::common_speculative_init_result(
const bool spec_mtp = std::find(params.speculative.types.begin(),
params.speculative.types.end(),
COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
GGML_ASSERT(has_draft || spec_mtp);

auto mparams = common_model_params_to_llama(params);
auto cparams = common_context_params_to_llama(params);
Expand Down Expand Up @@ -2560,34 +2521,6 @@ bool common_speculative_process(common_speculative * spec, const llama_batch & b
return result;
}

bool common_speculative_need_embd(common_speculative * spec) {
if (spec == nullptr) {
return false;
}

for (auto & impl : spec->impls) {
if (impl->need_embd()) {
return true;
}
}

return false;
}

bool common_speculative_need_embd_nextn(common_speculative * spec) {
if (spec == nullptr) {
return false;
}

for (auto & impl : spec->impls) {
if (impl->need_embd_nextn()) {
return true;
}
}

return false;
}

void common_speculative_draft(common_speculative * spec) {
if (spec == nullptr) {
return;
Expand Down Expand Up @@ -2672,7 +2605,10 @@ void common_speculative_draft(common_speculative * spec) {
void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, uint16_t n_accepted) {
common_speculative_impl * impl = spec->impl_last[seq_id];

GGML_ASSERT(impl);
if (impl == nullptr) {
GGML_ASSERT(n_accepted == 0);
return;
}

{
common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
Expand Down
6 changes: 0 additions & 6 deletions common/speculative.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,12 +67,6 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
// process the batch and update the internal state of the speculative context
bool common_speculative_process(common_speculative * spec, const llama_batch & batch);

// true if any implementation requires target post-norm embeddings to be extracted
bool common_speculative_need_embd(common_speculative * spec);

// true if any implementation requires target nextn embeddings to be extracted
bool common_speculative_need_embd_nextn(common_speculative * spec);

// generate drafts for the sequences specified with `common_speculative_get_draft_params`
void common_speculative_draft(common_speculative * spec);

Expand Down
47 changes: 42 additions & 5 deletions examples/speculative-simple/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,10 +3,47 @@
Demonstration of basic greedy speculative decoding

```bash
# spec-type draft-simple
./bin/llama-speculative-simple \
-m ../models/qwen2.5-32b-coder-instruct/ggml-model-q8_0.gguf \
-md ../models/qwen2.5-1.5b-coder-instruct/ggml-model-q4_0.gguf \
-f test.txt -c 0 -ngl 99 --color on \
--sampling-seq k --top-k 1 -fa on --temp 0.0 \
-ngld 99 --spec-draft-n-max 16 --spec-draft-n-draft-min 5 --draft-p-min 0.9
-hf ggml-org/Qwen3-8B-Base-GGUF:Q8_0 \
-hfd ggml-org/Qwen3-0.6B-Base-GGUF \
-p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
--spec-type draft-simple --spec-draft-n-max 7 -ngld 99 --color on \
-n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4

# spec-type draft-mtp
./bin/llama-speculative-simple \
-hf ggml-org/Qwen3.6-27B-GGUF:Q8_0 \
-p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
--spec-type draft-mtp --spec-draft-n-max 3 -ngld 99 --color on \
-n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4

# spec-type draft-mtp (with shared KV cache)
# note: this model needs a <s> token at the start to somewhat work without the chat template
./bin/llama-speculative-simple \
-hf ggml-org/Gemma-4-31B-it-GGUF:Q8_0 \
-p "<s>Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
--spec-type draft-mtp --spec-draft-n-max 3 -ngld 99 --color on \
-n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4

# spec-type draft-eagle3
./bin/llama-speculative-simple \
-hf ggml-org/gpt-oss-20b-GGUF \
-p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
--spec-type draft-eagle3 --spec-draft-n-max 3 -ngld 99 --color on \
-n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4

# spec-type draft-dflash
./bin/llama-speculative-simple \
-hf ggml-org/Qwen3-8B-GGUF \
-p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
--spec-type draft-dflash --spec-draft-n-max 7 -ngld 99 --color on \
-n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4

# spec-type draft-dspark
./bin/llama-speculative-simple \
-hf ggml-org/Qwen3-8B-GGUF \
-p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
--spec-type draft-dspark --spec-draft-n-max 7 -ngld 99 --color on \
-n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
```
Loading
Loading